Extracting Weights about nonlinear-gc HOT 4 OPEN

Hi Avinash,

You've noticed that the experiment files typically save a dictionary of results, with an entry called 'GC_est'. The way this 1D array is computed is: i) the relevant first layer weights are extracted from the neural network (all weights in MLP; weights that touch the input in LSTM), ii) those weights are reshaped if applicable (in MLP we reshape the weights to put all the lags for one input variable in the same column), iii) we take the norm of the weight matrix to quickly determine if there are columns where all weights are equal to zero.

If you follow a similar pattern in your experiments, you can determine the sparsity pattern by checking where zeros are located in the 'GC_est' entry in the results dictionary.

As for data preprocessing, results will definitely vary depending on the scaling of your inputs. It would probably be reasonable to standardize your input data (as people do in linear l1-regularized models) by scaling each feature to be in the range [-1, 1] or to have variance 1.

Let me know if there's anything else I can help clarify.

Ian

from nonlinear-gc.

Thanks,
I am trying to run a simple bivariate time series to check causality. I am getting the following error:
RuntimeError: size mismatch, m1: [49992 x 20], m2: [10 x 10] at /pytorch/aten/src/TH/generic/THTensorMath.cpp:940

I am guessing there is a layer somewhere in the mlp.py file which needs some sort of size specifications. Do you know how to fix this? I am a little knew to pytorch hence I am asking these many questions.

with thanks,
Avinash

from nonlinear-gc.

Hi Avinash, sorry for the slow reply. Please take a look at this repository, where I've written a much cleaner implementation of these methods. There are a couple demos, so I think you'll find this much easier to use.

from nonlinear-gc.

No problem Ian, I will go through the new repository, thanks.

from nonlinear-gc.